A Real Application: Detection and Recognition of Soccer Highlights
Among the many sports types, soccer is for sure one of the most relevant and worldwide diffused. In the following sections we report on our experience in the classification of soccer highlights, using an approach based on temporal logic models. The method has been
using several soccer videos containing a wide range of different video editing and camera motion styles, as produced by several different international broadcasters. Considering a variety of styles is of paramount importance in this field, as
the system lacks robustness. In fact,
produced by different directors display different styles in the length of the shots, in the number of
, in the editing effects.
We review hereafter previous work
to soccer videos. The work presented in  is limited to detection and tracking of both the ball and the players; the authors do not attempt to identify highlights. In , the authors rely on the fact that the playing field is always green for the purpose of extracting it. Successive detection of ball and players is limited to the field, described by a binary mask. To determine position of moving objects (ball and players) within the field, the central circle is first located, and a four-point homographic planar transformation is then performed, to map image points to the model of the playing field. Whenever the central circle is not present in the current frame, a mosaic image is used to extend the search context. In this latter case, the mosaicing transformation is combined with the homographic transformation. This appears to be a
expensive approach. In  a hierarchical E-R model that captures domain knowledge of soccer has been proposed. This scheme organizes basic actions as well as complex events (both
and interpreted), and uses a set of (nested) rules to tell whether a certain event takes place or not. The system relies on 3D data of position of players and ball, which are obtained from either microwave sensors or multiple video cameras. Despite the authors' claim that, unlike other systems, their own works on an exhaustive set of events, only little evidence of this is provided, as only a basic action
and a complex event
) are discussed. In  has been proposed the usage of panoramic (mosaic) images to present soccer highlights: moving objects and the ball are super-imposed on a background image featuring the playing field. Ball, players and goal posts are
. However, despite the title, only presentation of highlights is addressed, and no semantic analysis of relevant events is carried out.
Analysis of the
and Extracted Features
Inspection of tapes showed that
of videos use a main camera to follow the action of the game; since game action depends on the ball position, there exists a strong correlation between the movement of the ball and camera action. The main camera is positioned along one of the long sides of the playing field. In Figure 5.13 some typical scenes taken with the main camera are shown.
taken from the main camera.
Identification of the part of the playing field currently framed and camera action are among the most significant features that can be extracted from shots taken by the main camera; these features can be used to describe and identify relevant game events. Typical actions featured by the main camera are:
) tilt and
) zoom. Pan and tilt are used to move from a part of the playing field to another one, while zoom is used to change the framing of the subject.
Highlights that we have elected for thorough investigation are: i) forward launches,
i) shoots on goal,
v) penalty kicks, free kicks
to the goal box and corner kicks. These are typical highlights shown in TV news and magazine programs summarizing a match, even if these actions do not lead to the scoring of a goal. Moreover, penalty kicks and corners are often used to calculate statistics for a match. All of the above highlights are part of attack actions taking place in the goal box area, and are therefore
to goal actions. It must also be noted that for each of the above highlights there exist two different versions, one for each playing field side. The system that we present can discriminate each case.